Actually, RAGs can be, in my practice, an additional agent tool.... and there are a few agents RAG architectures where agents boost the quality of the KB data and retrieval steps in different ways.
The RAG vs Agents tradeoff looks different when inference cost approaches zero. DeepSeek just made its 75% price cut permanent, output tokens now $0.88/M, with the technical report flagging further cuts as Huawei chip efficiency improves. Chinese labs are not debating this tradeoff. Alibaba's new cloud portal launched with a single entry point, one agent-readable npm command, no product catalog. Their Zhenwu M890 chip ran a model autonomously for 35 hours, 432 kernel evaluations, 1158 tool calls, no human intervention. When the hardware is designed for 35-hour autonomous runs and inference costs keep dropping, the RAG tradeoff (cheap, predictable, one retrieval) shifts. Agent loops become the default architecture, not the premium option.
Love this piece! The real issue is that RAG is an information architecture problem, while agents are an accountability problem. With RAG, the question is mostly: did we fetch the right context and ground the answer? That is painful, but at least you can debug it like software.
With agents, the question becomes: who is responsible for a chain of decisions nobody fully specified, nobody fully observed, and nobody can easily replay when it fails?
Actually, RAGs can be, in my practice, an additional agent tool.... and there are a few agents RAG architectures where agents boost the quality of the KB data and retrieval steps in different ways.
Excellencies, distinguished leaders,
We stand at the threshold of a transformation
that will define not just economies
but the future of humanity itself.
Artificial Intelligence is no longer a distant innovation.
It is here.
It is accelerating.
And it is reshaping the foundations of our world.
Across governments, industries, and institutions,
AI is already influencing:
How we work
How we secure our nations
How we deliver healthcare and education
And how decisions are made at scale
But with this extraordinary power comes
an equally profound responsibility.
Because Artificial Intelligence is not just a tool.
It is a system-shaping force.
One that can:
Drive unprecedented economic growth
Solve complex global challenges
And expand human potential
But also one that can:
Amplify misinformation
Disrupt labor markets at scale
Deepen inequality
And introduce new risks to security and stability
This is the defining paradox of AI:
Its greatest strengths
are inseparable from its greatest risks.
And that is why this moment demands leadership.
Not fragmented.
Not reactive.
But coordinated, forward-looking, and global.
We must move beyond the question of whether to regulate AI
and focus on how to govern it wisely.
This requires action on several fronts:
1. Establish Global Principles
We need shared frameworks that ensure AI is:
Safe
Transparent
Accountable
And aligned with human values
2. Prevent a Fragmented AI Landscape
Competing standards and regulatory divides
risk creating instability and technological inequality.
We must avoid a world where AI development is geopolitically divided.
3. Protect People and Economies
We must prepare for workforce disruption
through reskilling, education, and inclusive growth strategies.
4. Safeguard Against Misuse
From cyber threats to autonomous systems,
AI must not become a source of uncontrolled risk.
5. Ensure Equitable Access
AI must not widen the gap between nations
it must help close it.
Excellencies,
The choices we make today will determine
whether AI becomes a force for progress
or a driver of division.
History has shown us that technology alone
does not shape the future.
Leadership does.
Cooperation does.
Vision does.
This is our moment
to ensure that Artificial Intelligence
remains firmly in service of humanity
not the other way around.
Let us act with urgency.
Let us act with responsibility.
And let us act together.
Thank you.
The RAG vs Agents tradeoff looks different when inference cost approaches zero. DeepSeek just made its 75% price cut permanent, output tokens now $0.88/M, with the technical report flagging further cuts as Huawei chip efficiency improves. Chinese labs are not debating this tradeoff. Alibaba's new cloud portal launched with a single entry point, one agent-readable npm command, no product catalog. Their Zhenwu M890 chip ran a model autonomously for 35 hours, 432 kernel evaluations, 1158 tool calls, no human intervention. When the hardware is designed for 35-hour autonomous runs and inference costs keep dropping, the RAG tradeoff (cheap, predictable, one retrieval) shifts. Agent loops become the default architecture, not the premium option.
Love this piece! The real issue is that RAG is an information architecture problem, while agents are an accountability problem. With RAG, the question is mostly: did we fetch the right context and ground the answer? That is painful, but at least you can debug it like software.
With agents, the question becomes: who is responsible for a chain of decisions nobody fully specified, nobody fully observed, and nobody can easily replay when it fails?
There’s no such thing as RAG vs Agentic tool use. The both are designed for different purposes, and mix perfectly well together when needed :)